(IJACSA) International Journal of Advanced Computer Science and Applications,

Vol. 1, No. 5, November 2010

A suitable segmentation methodology based on pixel

similarities for landmine detection
in IR images
Dr. G.Padmavathi1 ,
Department of Computer Science,
Avinashilingam University for
Women ,Coimbatore, India
ganapathi.padmavathi@gmail.
com
Dr. P. Subashini2 , Ms. M. Krishnaveni3
Department of Computer Science, Department of Computer Science,
Avinashilingam University for Avinashilingam University for
Women ,Coimbatore, India Women ,Coimbatore, India
mail.p.subashini@gmail.com krishnaveni.rd@gmail.com

Abstract— Identification of masked objects especially in detection
of landmines is always a difficult problem due to environmental
inference. Here, segmentation phase is highly concentrated by
performing an initial spatial segmentation to achieve a minimal
number of segmented regions while preserving the homogeneity
criteria of each region. This paper aims in evaluating similarities
based segmentation methods to compose the partition of objects in
Infra-Red images. The output is a set of non-overlapping
homogenous regions that compose the pixels of the image. These
extracted regions are used as the initial data structure in feature
extraction process. Experimental results conclude that h-maxima
transformation provides better results for landmine detection by
taking the advantage of the threshold. The relative performance of
different conventional methods and proposed method are
evaluated and compared using the Global Consistency Error and
Structural Content. It proves that h-maxima gives significant
results that definitely facilitate the landmine classification system
more effectively.
Keywords- Segmentation, Global Consistency error, h-maxima,
threshold, Landmine detection
I. INTRODUCTION
The signature of buried land mine in IR images varies
significantly depending on external parameters such as
weather, soil moisture, solar radiation, burial depth, and
time[7]. By literature [Oscar González Merino] [11] the
working of many image-based landmine detection algorithms,
it is concluded that the fundamental challenges arise from the
fact that the mean spectral signatures of disturbed soil areas
that indicate mine presence are nearly always very similar to
the signatures of mixed background pixels that naturally occur
in heterogeneous scenes composed of various types of soil and
vegetation [1][12]. Mine detection using infrared techniques is
primarily based on exploiting temperature differences between
pixels on the mines and background pixels [2]. Thus, there is
always a need for robust algorithm that has the capability to
analyze the pattern of distribution of the pixels to separate
pixels of the mines from background pixels. Here the
segmentation methods are evaluated with performance metrics
[3]. Experimental results show that h-maxima transformation is
more adoptable for IR target images. The paper is organized as
follows: Section 2 deals with the need for pixel based
segmentation for IR images. Section 3 converses the
comparison of segmentation methods and the subjective
assessment of the approach Section 4 explores the performance
evaluation of the methods. The paper ends with observations
on future work and some conclusions.
II. PRELIMINARIES OF SEGMENTATION
There are many unsupervised and supervised segmentation
algorithms [6]. They only use low-level features, e.g. intensity
and texture, to generate homogeneous patches from an input
image. Four categories for segmentation are: histogram shape
based methods, where, for example, the peaks, valleys and
curvatures of the smoothed histogram[4]. Clustering based
methods, where the gray level samples are clustered in two
parts as background and foreground (object), or alternately are
modeled as a mixture of two Gaussians. Entropy based
methods results in algorithms that use the entropy of the
foreground and background regions, the cross entropy between
the original and binarized image, etc. Object attribute-based
methods search a measure of similarity between the gray level
and the binarized images, such as fuzzy shape similarity, edge
coincide, etc. This paper quantifies the segmentation methods
based on the similarities of the pixels including the intensity
and the object structure.
III. SIMILARITIES BASED SEGMENTATION TECHNIQUES
Segmentation is a pre-process which partitioned image into
unique multiple regions, where region is set of pixels.
Mathematically segmentation can be defined as follows:
If I is set of all image pixels, then by applying segmentation
we get different unique regions like {S1, S2, S3,…,Sn } which
when combined formed „I‟ . Basic formulation is as follows:
n
(a) Si  I where Si Sj  
i 1, n
(b) Si is a connected region, i=1, 2….n.
(c) P(S i) = TRUE for i=1, 2... n.
(d) P Si  S j   FALSE fori  j.

88 | P a g e
http://ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Where P (Si) is a logical predicate defined over the points
in set Si.
Condition (a) indicates that segmentation must be complete,
every pixel in the image must be covered by segmented
regions. Segmented regions must be disjoint. Condition (b)
requires that points in a region be connected in some
predefined sense like 4- or 8- connected. Condition (c) deals,
the properties must be satisfied by the pixels in a segmented
region- e.g. P(Si) = TRUE. if all pixels in Si have the same gray
level. Last condition (d) indicates that adjacent regions Si and
Sj are different in the sense of predicate P.
Ever in image processing research there is no common
solution to the segmentation problem [6]. One of the main
reasons of segmentation algorithms is to precisely segment the
image without under or over segmentation. Almost all image
segmentation techniques proposed so far are ad hoc in nature.
These below are the following approaches of image
segmentation taken in this paper and demonstrated with IR
images. Given below in fig 1 is the approach taken for image
segmentation for IR images.
Figure 1: Schematic representation of proposed algorithm for identification of
landmine
A. H-maxima algorithm
Every IR image taken has been implemented with
morphological reconstruction, extended maxima transformation
using thresholding[10] . The extended maxima transformation
is the regional maxima computation of the corresponding h-
maxima transformation. As a result, it produces a binary image.
A connected-component labeling operation is performed, in
order to evaluate the characteristics and the location of every
object. As a second object reduction step, objects not located
within a region of another object, are also discarded, since mine
objects are not typically clustered. The region of interest (ie)
the target is got by connected component segmentation in
which the relevant pixels of the object will be grouped and
http://ijacsa.thesai.org
Vol. 1, No. 5, November 2010
extorted [8]. The extended-maxima transform computes the
regional maxima of the H-maxima transform. Here H refers to
nonnegative scalar. Regional maxima are connected
components of pixels with a constant intensity value, and
whose external boundary pixels will have a lower value.
B. Kmeans algorithm
Here K means is used as a two phase iterative algorithm[9]
to minimize the sum of point-to centroid distances, summed
over all k clusters. The first phase uses each iteration that
consists of reassigning points to their nearest cluster centroid.
The second phase uses points that are individually reassigned.
Arbitrarily choose k data points to act as cluster centers. Until
the cluster centers are unchanged following are the steps
carried out : Allocate each data -point to cluster whose center is
nearest. Replace the cluster centers with the mean of the
elements in their clusters end. Clustering in attribute space can
lead to unconnected regions in image space (but this may be
useful for handling occlusions).The K means image
representation groups all feature vectors from all images into K
clusters and provides a cluster id for every region of image that
represents the salient properties of the region. K means is fast
iterative and leads to a local minimum. It looks for unusual
reduction in variance. This iterative algorithm has two steps
Assignment step: Assign each observation to the cluster with
the closest mean
Si(t )  { X j : X j  mi(t )  x j  m( t ) --------- (1)
Update step: Calculate the new means to be centroid of the
observations in the cluster
1
mi( t 1) 
Si( t )
X j -
x j S ( t ) i
-(2)
C. Threshold intensity distribution algorithm
It can detect object boundaries with low gradient or
reduce noise effect in gradient. However, an accurate and
stable estimation of intensity distribution is difficult to get
from a finite set of 3D image data. To reduce the ”shrink” or
”expand” effect on segmentation results, gradient information
is used to calibrate the estimation of intensity distribution in
the following. Overlap of image gradient is computed with
the boundaries determined by intensity distribution through
introducing a probability offset to intensity distribution. The
maximum overlap indicates the optimal boundaries of the
interested objects. To restate the problem without losing
generality, here it use a mixed Gaussian distribution model.
n
P(u )    k P(u k ; k , k ),
k 1
89 | P a g e
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 1, No. 5, November 2010

Where k is the prior probability of class k with

pixels belong to an object and 0 pixels constitute the
background.

k ,  k are the mean and variance of the

n


k 1
k  1, and Rules that are worked out to guide the process of true
corner point localization.
Gaussian distribution of the intensity. Intensity distribution (i) Select those boundary points which bear significantly large
inside the region Ω is in Eq below cornerity index by eliminating the boundary points which lie
on straight line segments bearing negligibly small cornerity

  P u  ;  , 
Pin (u )  k k k k
k 
k 
And the intensity distribution of the outside region
Ω, Pout , can be obtained in a similar way. Normally for pixel
x on region boundaries with u = I(x),
Pout  I ( x)   Pin  I ( x)   0
Threshold Intensity Distribution is a significant aspect in
assigning an individual region around an image[4]. Uneven
distribution may lead to assignment of two or more regions to
an individual pixel. This method reduces processing time by
performing gray-level based segmentation that extracts regions
of uniform intensity. Subsequently, it is also possible to
estimate motion for the regions. It also reduces the
computational load, and the region-based estimator gives
robustness to noise and changes of illumination. The
segmentation of the reference image is designed to group pixels
of similar gray-levels. In intensity distribution, B (i,j) is a
binary image (pixel are either 0 or 1) created by thresholding
F(i,j)
B(i,j) = 1 if F (i,j)<t
B (i,j) =0 if F(I,j)>=t
index value.
(ii) Since all points on a smooth curve segment are in general
associated with almost same cornerity index, and actual corner
points bear cornerity index larger than that of their neighbors,
it is suggested to select the set of connected points such that
the variations in their cornerity indices are considerably large.
This rule helps in selecting only the set of points with in the
vicinity of actual corner points by eliminating the points on
smooth curve segments.
(iii) Select the points which bear local maximum cornerity
index as true corner points. It could be noticed that these rules
do not require any priori knowledge in locating true corner
points.
Thus, the expected point corresponding to a corner point
will have a larger shift when compared to other points on the
boundary curve. Therefore, the cornerity index of pi is defined
to be the Euclidean distance d between the points pi and its
expected point pie and is given by
d  xi  xie    yi  yie 
2 2
The cornerity index indicates the prominence of a corner
point. The larger the value of the cornerity index of a boundary
point, the stronger is the evidence that the boundary point is a
corner.

It is assumed that the 1‟s are the object pixels and the 0‟s
are the background pixels.
The Histogram (h) - gray level frequency distribution of the
gray level image F.

hF (g) = number of pixels in F whose gray level is g

HF (g) =number of pixels in F whose gray level is <=g

This method is a probabilistic method that makes

parametric assumptions about object and background intensity
distributions and then derives “optimal” thresholds.

D. Boundary based algorithm

Here the strategy is to consider each control point in turn
and move it to the pixel; in its local neighbourhood which gives
us the minimum. For a closed boundary it could make the
initial estimate surround the object of interest, and add in
another term to the objective function to penalize the total
length. A difficulty with this type of strategy is the control
points. This method traces the exterior boundaries of objects, as
well as boundaries of holes inside these objects, in the binary
image. A binary image is considered in which the nonzero

90 | P a g e
http://ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 1, No. 5, November 2010

Global Consistency Error (GCE) forces all local

refinements to be in the same direction and is defined as:
GCE (S , S ')  min  LRE (S , S ', xi ),  LRE (S ', S , xi ) .(3)
1
N

Global Consistency Error

1.2

Range of values
1 boundary
0.8
intensity
0.6
kmeans
0.4
0.2 h-transform
0

e1

e2

e3

e4

e5

e6

e7

e8

e9

0
e1
ag

ag

ag

ag

ag

ag

ag

ag

ag
ag
Im

Im

Im

Im

Im

Im

Im

Im

Im
Im
Images taken

Figure 3: Comparison based on Global Consistency Error

Structural Content

35
Range of values

30
boundary
25
20 intensity
15 kmeans
10
h transform
5
0
e1

e2

e3

e4

e5

e6

e7

e8

e9

0
e1
ag

ag

ag

ag

ag

ag

ag

ag

ag

ag
Im

Im

Im

Im

Im

Im

Im

Im

Im

Im
Images Taken

Figure 2: Visual assessment of the segmentation methods Figure 4: Comparison based on structural content

A structural content is used to find mask structural defects

IV. PERFORMANCE EVALUATION
and noise, but are prone to periodic artifacts. This examines to
The qualitative and quantitative assessment of segmentation classify image regions hierarchically based on the level of
results is carried out here for choosing the appropriate approach structural content.
for a given segmentation task. Similar to the segmentation
theory itself, there is no established standard procedure for the
M N M N
SC   x j , k 2 2 /  xj , k 2 2 / ……….(4)
evaluation of its results. For this reason evaluation is done
using empirical discrepancy method using the relative ultimate
measurement accuracy. Global Consistency Error and j 1 k 1 j 1 k 1
Structural Content used as the evaluation parameters to control
the segmentation process and dynamically a good number of Structural content of a segmented image is calculated by
regions are chosen based on local minima in the segmentation summation of the original image by the division of the
evaluation measure[5] .The uniqueness of each parameter takes segmented image.
the advantage and disadvantages of IR images.

91 | P a g e
http://ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
From the above figures 3, 4 the objective evaluation also
states that h maxima transform is the desirable method for
landmine detection. The Global consistency error rate should
always be low and it is the same for the proposed method in
comparatively to the conventional methods. The structural
content is high for h-maxima which state that the originality of
the object and its information remains the same after the
segmentation process.
V. CONCLUSION
The method described in this paper provides a relatively
simple, extremely fast, and robust method for displaying and
performing automatic target identification phase. The result of
image segmentation is a set of segments that collectively cover
the entire image, or a set of contours extracted from the image.
Each of the pixels in a region is similar with respect to some
characteristic or computed property, such as color, intensity, or
texture. Adjacent regions are significantly different with
respect to the same characteristic(s). From this paper it is
concluded that h-maxima is very useful in IR image evaluation
but that it should probably restrict studies to similar images and
similar processing. Simulations are carried out which
demonstrates the proposed method is able to successfully
localize landmine objects from different sets of real IR images.
Nevertheless, this scheme has some limitations because it is not
automatic as different parameters have to be adjusted manually.
Future work should incorporate the use of high-level image
analysis methods for the identification of the true mine objects
among the set of the detected mine cues.
REFERENCES
[1] Frank cremer't, wim de jong and klamer schutte ,” processing of
polarimetric infrared images for landmine detection”,2nd international
workshop on advanced gpr. 14-16 may, 2003, the Netherlands
[2] E Ctrmn, W. de long, and K Schutte, “Fusion of polarimetric infrared
features and GPR features far landmine detection,” in 2nd International
Workshop on Advanced Ground Penetrating Radar (IWAGPR), Delft,
The Netherlands, May 2003.
[3] Nathir A. Rawashdeh, Shaun T. Love ” Hierarchical Image
Segmentation by Structural Content” journal of software, vol. 3, no. 2,
february 2008
[4] Yongsheng Hu, Qian Chen,”An Infrared Image Preprocessing Method
Using Improved Adaptive Neighborhoods “,Proceedings of the 6th
World Congress on Intelligent Control and Automation, June 21 - 23,
2006, Dalian, China
[5] Zhou Wang, Alan Conard Bovik, Hamid Rahim Sheik and Erno P
Simoncelli, “Image Quality Assessment: From Error Visibility to
Structural Similarity”, IEEE Trans. Image Processing, and Vol. 13,
(2004).
[6] R.M. Haralick, and L.G. Shapiro, “Survey: Image Segmentation
Techniques,” Computer Vision, Graphics, and Image Processing, Vol.
29, pp. 100-132, 1985.
http://ijacsa.thesai.org
Vol. 1, No. 5, November 2010
[7] F. Cremer and W. de Jong and K. Schutte.„Infrared polarization
measurements of surface and buried anti-personnel landmines‟. In
Preprint Proc. SPIE Vol. 4394, Detection and Remediation Techniques
for Mines and Minelike Targets VI, Orlando FL, USA, Apr. 2001.
[8] Yan Sun, Chengyi Sun and Wanzhen Wang, “Color Images
Segmentation Using New Definition of Connected Components”,
Proceedings of ICSP2000.
[9] K.S. Ravichandran,B. Ananthi, “Color Skin Segmentation Using K-
Means Cluster” International Journal of Computational and Applied
Mathematics ISSN 1819-4966 Volume 4 Number 2 (2009), pp. 153–157
[10] Punam K. Saha, “Optimum Image Thresholding via Class Uncertainty
and Region Homogeneity”,IEEE transactions on pattern analysis and
machine intelligence, vol. 23, no. 7, july 2001
[11] Oscar González Merino “Image analysis of Infrared Polarization
measurements of Landmines”, Academic year: 2000/01
[12] H. Brunzell, “Clutter Reduction and Object Detection in Surface
Penetrating Radar,” in Proc. Radar.97 Conf., 1997 pp. 688-691.
AUTHORS PROFILE
Dr. Padmavathi Ganapathi is the Professor and
Head of Department of Computer Science,
Avinashilingam University for Women,
Coimbatore. She has 23 years of teaching
experience and one year Industrial experience.
Her areas of interest include Network security and
Cryptography and real time communication. She
has more than 108 publications at national and
International level. She is a life member of many
professional organizations like CSI, ISTE, AACE,
WSEAS, ISCA, and UWA.
Dr. Subashini is the Associate professor in
Department of Computer Science, Avinashilingam
Deemed University for Women, Coimbatore. She
has 16 years of teaching experience. Her areas of
interest include Object oriented technology, Data
mining, Image processing, Pattern recognition.
She has 55 publications at national and
International level.
Ms.M.Krishnaveni has 4 Years of Research
Experience Working as Research staff in DRDO
and pursuing her Ph.D in Avinashilingam
University for Women, Coimbatore. Her research
interest are Image Processing, Pattern
Recognition and Neural Networks.She has 24
publications at national and international level.
92 | P a g e